NVIDIA Debuts Agent Toolkit And NemoClaw At GTC For Faster, Safer AI Agents
The open-source NVIDIA Agent Toolkit is a developer platform that’s designed to standardize and help scale autonomous AI agents across enterprises. And NemoClaw is a secure software stack that layers privacy and runtime controls on top of OpenClaw personal agents, to address some of the concerns regarding the tool since its release late last year. Both are meant to accelerate agentic AI deployments and are optimized for NVIDIA’s many platforms, from the diminutive DGX Spark to big-iron Vera Rubin servers.
The NVIDIA Agent Toolkit And NemoClaw Ease Deployment And Increase Security
The NVIDIA Agent Toolkit is essentially a developer platform designed for building, orchestrating, and deploying autonomous AI agents at scale. NVIDIA mentioned a myriad of partners that have already committed to building with the Agent Toolkit, including Adobe, Salesforce, SAP, ServiceNow, Siemens, CrowdStrike, Atlassian, and Palantir, among others. According to NVIDIA, the Agent Toolkit offers faster integration of agentic AI capabilities into business applications, with shared resources and standards aimed at mitigating duplicated engineering efforts, and tight coupling with NVIDIA’s compute stack for optimized performance.
As its name suggests, NemoClaw augments the wildly popular OpenClaw, to enable easier deployment and also incorporate privacy and security guardrails. NemoClaw is a pre-packaged software stack that installs OpenClaw, along with NVIDIA’s Nemotron open models and the new OpenShell runtime with a single command, adding privacy, security, and operational controls to OpenClaw agents – dubbed Claws -- so they can run more safely and securely. NemoClaw is designed for quick setup across environments including everything from NVIDIA RTX-powered PCs, the DGX Station, DGX Spark, and NVIDIA-powered servers.
NVIDIA is releasing NemoClaw in an attempt to address some of the risks associated with self-evolving personal agents. NemoClaw and the OpenShell runtime add sandboxing, privacy controls, policy-based security controls, and enterprise-grade governance layers on top of OpenClaw agents.
DGX Spark And DGX Station For NemoClaw
NVIDIA also announced some updates regarding its DGX Spark and DGX Station development platforms. If you recall, the tiny DGX Spark is powered by the Grace Blackwell-based GB10 chip, which delivers up to 1,000 TOPS of AI compute power, and features NVIDIA's on-package NVLink C2C interconnect, which provides five times the bandwidth of a PCIe x16 slot between its CPU and GPU resources. This allows the CPU and GPU to share a single memory pool, and helps keep both better utilized with various workloads.
The DGX Station is at the opposite end of the spectrum. The DGX Station is built on the GB300 Grace Blackwell Ultra Superchip, which cranks compute performance up to 20 petaflops, or 20,000 TOPS. NVIDIA outfits the DGX Station with a whopping 768 GB of memory which should be plenty to hold virtually any model for the foreseeable future. On top of that, NVIDIA also packs each DGX Station with its ConnectX-8 SuperNIC, which makes these systems scale with 800 Gigabits per second of network bandwidth. This allows multiple DGX Stations to be networked together and distribute the load at very high-speeds.
Pairing NVIDIA DGX Spark and NVIDIA DGX Station systems with the new NVIDIA NemoClaw open source stack provides optimized platforms for locally developing and deploying Claws. NVIDIA also announced that DGX Spark now supports clustering up to four systems in a unified configuration (versus two at launch), to create a powerful, compact compute cluster with up to linear performance. An upcoming software update for DGX Spark will further strengthen orchestration and manageability, as well.
New Models And The Nemotron Coalition
There were a multitude of additional model-related announcements also made at the event—too many for us to cover them all here. NVIDIA announced a new partnership with Adobe to accelerate the company’s work in AI. Adobe will be building agents on NVIDIA’s Agent Toolkit, it’ll integrate Cuda-X and Omniverse libraries into its applications, and collaborate to build next-gen Firefly models trained on NVIDIA’s AI infrastructure, optimized for creative workflows. NVIDIA announced an array of physical AI and autonomous vehicle updates too, including the launch of NVIDIA Halos OS, and Alpamayo 1.5, which first debuted at CES earlier this year. Alpamayo is NVIDIA’s open-source family of Vision-Language-Action (VLA) models for reasoning-based decision making in future autonomous vehicles (AVs).
The Nemotron Coalition was also announced at GTC 2026. The Nemotron Coalition brings together NVIDIA, Black Forest Labs, Cursor, LangChain, Mistral AI, Perplexity, Reflection AI, Sarvam and Thinking Machines Lab, to combine their collective expertise to collaboratively build, next-generation open frontier models. The coalition will use NVIDIA’s internal DGX Cloud compute resources for model training, so no one has to duplicate the effort on training base models. The Nemotron Coalition is tasked with building shared, open foundations, that developers can then specialize for their respective industries, regions and use cases. The efforts of the coalition will form the foundation of the upcoming Nemotron 4 family of models.
